Five architectures. One workflow.
Unified, MoT, MoT-Joint, MoT-IDM and Hidden share data, training, checkpoint and evaluation interfaces.
WORLD ACTION MODELSOPEN RESEARCH
A unified research codebase
EasyWAM: A Unified and Efficient Framework for Training and Evaluating World Action Models
EasyWAM connects modular World Action Models, data processing, distributed training, parameter-efficient fine-tuning and scalable evaluation in one reproducible workflow.
01 / overview
One framework.
Every layer in sync.
From sparse video decoding to resumable evaluation, EasyWAM makes the entire WAM loop inspectable and repeatable.
Three video backbones.
Wan2.2-TI2V-5B, Cosmos-Predict2.5-2B and FLUX.2 Klein-4B, with full-parameter training and LoRA fine-tuning.
Efficient from data to inference.
FlashAttention 2/3/4, sparse video decoding, text and prompt caches, DeepSpeed ZeRO, dynamic inference batches and resumable evaluation.
Five benchmark workflows.
LIBERO, LIBERO-Plus, RoboTwin, RoboDojo and RoboCasa365, with distributed evaluation and result summaries.
02 / quick start
From setup
to evaluation.
Choose a task and copy commands from the project README. Follow the linked guides to prepare datasets, model weights and simulators.
01 / ENVIRONMENT
Full setup guide ↗Install EasyWAM
Run from the repository root after cloning. FlashAttention is optional and depends on your GPU.
TERMINAL / BASH
conda create -n easywam python=3.10 -y
conda activate easywam
pip install -U pip
pip install torch==2.7.1 torchvision==0.22.1 --extra-index-url https://download.pytorch.org/whl/cu128
pip install -e .02 / TRAINING
Training configuration ↗Precompute and train
After preparing LIBERO data and the backbone, precompute text embeddings and launch the example MoT run. Adjust the GPU process count for your machine.
TERMINAL / BASH
python scripts/precompute_text_embeds.py task=libero_easywam_mot_wan22
NPROC_PER_NODE=8 bash scripts/train_zero1.sh task=libero_easywam_mot_wan2203 / EVALUATION
Benchmark setup ↗Choose a benchmark
Set up the simulator and data with its guide, then replace the checkpoint placeholder with your own file.
TERMINAL / BASH
python experiments/libero/run_libero_manager.py \
task=libero_easywam_mot_wan22 \
ckpt=<path/to/checkpoint.pt>03 / results
Measured results,
across backbones.
Task success rates (%), higher is better. Results may vary with checkpoints, seeds and evaluation settings.
EasyWAM-Hidden average success
EasyWAM-MoT average success
EasyWAM-Unified average success
Explore all five result tables
| Model | Spatial | Object | Goal | LIBERO-10 | Avg. |
|---|---|---|---|---|---|
| EasyWAM-Unified | 98.4 | 98.8 | 99.2 | 98.0 | 98.6 |
| EasyWAM-MoT | 97.0 | 99.2 | 96.6 | 94.0 | 96.7 |
| EasyWAM-Hidden | 99.2 | 100.0 | 97.8 | 98.2 | 98.8 |
| Model | Spatial | Object | Goal | LIBERO-10 | Avg. |
|---|---|---|---|---|---|
| EasyWAM-Unified | 91.2 | 98.8 | 91.8 | 66.2 | 87.0 |
| EasyWAM-MoT | 96.8 | 99.6 | 97.4 | 90.0 | 95.9 |
| EasyWAM-Hidden | 98.0 | 99.8 | 89.4 | 86.6 | 93.5 |
| Model | Spatial | Object | Goal | LIBERO-10 | Avg. |
|---|---|---|---|---|---|
| EasyWAM-Unified | 97.8 | 99.4 | 97.0 | 93.0 | 96.8 |
| EasyWAM-MoT | 98.0 | 98.4 | 98.4 | 95.6 | 97.6 |
| EasyWAM-Hidden | 97.2 | 98.8 | 96.0 | 95.0 | 96.8 |
| Model | Orig | Background | Camera | Language | Layout | Light | Noise | Robot | Avg. |
|---|---|---|---|---|---|---|---|---|---|
| EasyWAM-Unified | 98.6 | 72.3 | 54.8 | 93.7 | 83.4 | 97.0 | 72.0 | 83.4 | 79.0 |
| EasyWAM-MoT | 96.7 | 64.5 | 45.5 | 71.4 | 80.1 | 94.7 | 78.5 | 71.5 | 71.7 |
| EasyWAM-Hidden | 98.8 | 59.3 | 57.0 | 93.2 | 84.3 | 95.0 | 70.8 | 83.7 | 77.6 |
| Model | Orig | Background | Camera | Language | Layout | Light | Noise | Robot | Avg. |
|---|---|---|---|---|---|---|---|---|---|
| EasyWAM-Unified | 96.8 | 72.7 | 63.7 | 90.1 | 82.6 | 89.2 | 72.1 | 79.2 | 78.2 |
| EasyWAM-MoT | 97.6 | 60.7 | 75.1 | 92.3 | 82.6 | 92.6 | 81.3 | 58.5 | 77.7 |
| EasyWAM-Hidden | 96.8 | 59.5 | 65.9 | 92.6 | 85.0 | 89.1 | 68.3 | 82.5 | 77.8 |
LIBERO-Plus uses LIBERO checkpoints and tests changes to background, camera, language, layout, lighting, noise and robot appearance.
04 / efficiency
Less repeated work,
faster iteration.
EasyWAM provides efficiency controls across attention, memory, data loading and batched evaluation.
Efficient attention
Choose the fastest compatible FlashAttention 4, 3 or 2 backend, with PyTorch SDPA as a fallback. BF16 training reduces tensor bandwidth and memory use.
Memory controls
LoRA, DeepSpeed ZeRO, gradient checkpointing and VAE micro-batching help fit larger workloads, with configurable speed and memory tradeoffs.
Lean data pipeline
Sparse video decoding, indexed text caches, precomputed text embeddings and persistent DataLoader workers reduce repeated loading and processing.
Batched evaluation
Dynamic inference batching, multiple model workers per GPU, prompt caching and resumable evaluation improve GPU use and avoid repeated work.
FastWAM: 51.5 samples/s
Versus the original FastWAM codebase
About 14 hours with original FastWAM
These results use the Wan2.2-TI2V-5B backbone on 8 × NVIDIA H100 GPUs with per-device batch size 16. EasyWAM-MoT and FastWAM share the same model architecture. Actual training time varies with CPU, GPU and data configuration.
| Framework | Throughput (samples/s) ↑ | Data (s) ↓ | Forward (s) ↓ | Backward (s) ↓ |
|---|---|---|---|---|
| EasyWAM-MoT | 138.2 | 0.0108 | 0.3663 | 0.5392 |
| FastWAM | 51.5 | 0.0759 | 1.2412 | 1.0680 |
05 / model zoo
Multiple models,
shared interfaces.
These architectures and backbones are supported by the framework. Check the Hugging Face collection for released checkpoints.
EasyWAM-Unified
Jointly predicts future video and actions in a single Video DiT.
EasyWAM-MoT
Video DiT and Action DiT interact through mixed self-attention for action-only prediction.
EasyWAM-MoT-Joint
Jointly denoises video and action tokens through mixed self-attention.
EasyWAM-MoT-IDM
Uses teacher-forced conditional video for action prediction.
EasyWAM-Hidden
Conditions a separate Action DiT on intermediate Video DiT features.
SUPPORTED BACKBONES
Browse released models ↗Wan2.2-TI2V-5B · Cosmos-Predict2.5-2B · FLUX.2 Klein-4B
06 / community
Explore further,
build together.
Find guides, models and updates. Bring questions and ideas to the GitHub community.
01 / DOCS
Documentation
Guides for backbones, data, training configuration and benchmarks.
Read the docs ↗ 02 / CODEGitHub
Browse code, configurations, updates and contribution guidance.
View repository ↗ 03 / MODELSHugging Face
Find published EasyWAM checkpoints and datasets.
Browse collection ↗ 04 / FEEDBACKFeedback and discussion
Report issues, propose features or join the project discussion.
Open issues ↗WECHAT COMMUNITY
Stay in touch with EasyWAM
Scan to join the WeChat discussion group, or email the project team.
Email the team ↗